Investigate sqlite performance for large tables.

[evetion]

Since we're dropping arrow, and netcdf might not be optimal for some tables. Sqlite is used for pretty large databases, so our geopackage/python code could be optimized by some pragmas further still.

#2521

[Huite]

netcdf might not be optimal for some tables

Which tables are under consideration?

[evetion]

Whereas basin time seems a natural fit for netcdf, the three directly linked control tables (also with their truth states) feel a bit ugly.

As replacement for arrow, netcdf should be able to fit all tables, except for the geometry. If we could fix the gpkg performance, we might not need it for input.

[evetion]

It seems SQLite.jl is a bit slow. It takes 25 s on a large Basin / time table, whereas the same can be done in geopandas in 15 s. Hopefully we can get that down even further.

edit: JuliaDatabases/SQLite.jl#346 can get it to run in 5 s.